Vehicle Routing to Minimize Mixed-Fleet Fuel Consumption and Environmental Impact

O.Gusikhin1, P. MacNeille1, A. Cohn2,

1 Ford Motor Company Research and Innovation Center 2101 Village Road, Dearborn, MI 48121, U.S.A 2 University of Michigan

Abstract: Efficient vehicle routing is critical to the operational profitability and customer satisfaction of vehicle fleet-related businesses, especially in light of increasing, and highly volatile, fuel prices. Growing pressures to reduce negative environmental impacts have suggested that a second metric (vehicle emissions) should also be considered in vehicle routing. Currently, the majority of existing tools use distance as a surrogate for cost. When considering a mixed fleet of multiple vehicle types, with individual vehicles within a fleet type also varying by age and vehicle health, this surrogate becomes significantly less accurate. Furthermore, using distance as a surrogate fails to capture the variations between city and highway driving, which are particularly striking for hybrid vehicles. We thus propose a new approach to the vehicle routing problem, specifically targeting applications with mixed fleets including cleanvehicle technologies, in recognition of the limitations of the existing approaches.

1 Introduction
Efficient vehicle routing is critical to the operational profitability and customer satisfaction of vehicle fleet-related businesses. The rising and highly-variable cost of fuel, as highlighted by the price spike in the summer of 2008, increases the importance of efficient vehicle routing. At the same time, growing environmental concerns suggest that cost is not the only metric of importance, and that emissions should also be taken into account. Currently, the majority of existing methods and software packages minimize travel distance as a surrogate for cost. Although there is a positive correlation between distance traveled and fuel consumed, it is not a perfect correlation. In particular, both fuel efficiency and emissions vary depending upon driving conditions (e.g. city vs. highway driving). This variability is even more pronounced in a heterogeneous fleet comprised of multiple vehicle types with a range of fuel-consumption and emissions characteristics. In recent years, socio-economical pressures to reduce their fuel costs and CO2 footprint have motivated many fleet operators to begin upgrading their fleets, focusing on clean-vehicle technologies and alternative fuels, including flex- fuel,

plug-in electric vehicles (PEV), hybrid vehicles, and plug-in hybrid electric vehicles (PHEV). As a result, many commercial fleets are currently composed of a heterogeneous set of vehicles, with noticeable variations in fuel economy and emissions across vehicle type. Furthermore, there are variations across vehicles even within a given vehicle type or common set of capabilities, because newer vehicles typically exhibit better fuel economy and better emission control technologies than older vehicles. Consider the following examples of highly heterogeneous fleets: Florida Power and Lighting (www.fpl.com), the leader in green fleet initiatives, has a fleet of approximately 2400 vehicles, with half of the fleet powered by biodiesel, 300 hybrids and plug-in hybrids now in service, and plans to convert one-third of the vehicles to hybrid by the end of 2010. A key issue in incorporating the fuel efficiencies and emissions of heterogeneous fleets within the vehicle routing problem is this: The differences in fuel efficiencies and emissions across vehicle types are not exclusively proportional to the distance traveled, but are also highly dependant on the driving cycle. For instance, a hybrid vehicle takes advantage of the regenerative breaking that occurs in stop-and-go driving environments to charge a battery which can then be used to power the vehicle when traveling at slow speeds. On the average fuel consumption is reduced about 20% [1] by regenerative braking. Conversely, in highway driving, hybrids provide minimal, if any, improvement in fuel efficiencies over their conventional counterparts. The introduction of PHEVs, which use electrically-charged batteries for the initial portion of a trip and revert to gasoline-based power only when the battery has been depleted, presents even greater variability in terms of the correlations between distance and fuel utilization. When we move beyond fuel costs to also consider the environmental impact of vehicle routing, the complexity grows further. Fuel consumption is effectively proportional to carbon dioxide (CO2) emission, so fuel economy improvements are reflected in CO2 reduction. Additionally, there may be requirements to minimize or eliminate conventional emissions such as volatile organic components (VOC) and nitrogen oxides (NOx) emissions in highly populated areas, requiring the use of electric power or other clean alternative fuel options and thus limiting the feasible region of a vehicle routing problem. In this paper, we consider ways to explicitly capturing fuel consumption and emissions in a mixed-fleet vehicle routing program and analyze the opportunities for simultaneously reducing costs and negative environmental impacts. In Section 2, we review factors that influence fuel consumption and emissions, and consider ways to estimate these metrics for a given route. In Section 3, we suggest a number of formulations for this new variation of the vehicle routing problem. We also outline a solution approach based on composite variable modeling. In Section 4, we provide a numerical example and analysis to highlight the benefits of our proposed approach and we then offer conclusions and suggested areas for future research in Section 5.

2 Estimation of Fuel Economy & Vehicle Environmental Impact (factors that influence fuel consumption and emissions)
Route planning is done by representing the road system as a graph in which intersections are nodes and road segments are arcs. To determine the best route from an origin node to a destination node, each arc is associated with a cost that represents distance, travel time, or fuel consumption. Then Dijkstra's algorithm (or an equivalent) is used to find the lowest cost path which is then inversed mapped to the road system for visualization and navigation of the preferred route. Total fuel cost along a given route is the product of the total consumption of each type of fuel and the per-unit cost for that fuel. Total fuel consumption along an arc is largely dependant on the vehicle specific load (VSL) opposing vehicle motion multiplied by the distance traveled against the VSL to give the total energy required to travel the arc. This energy can be readily converted into fuel, for example about 0.0003 gallons of gasoline or 0.0002 kWh of electricity are needed to push against a pound of force over a one mile stretch of road. These factors will vary, however; with the efficiency of the energy conversion which depends on many factors. The VSL depends largely on external factors that include the drive cycle, aerodynamic drag, rolling resistance, parasitic drag, and gradient drag. These factors in turn are dependant on several external factors including weather conditions, road and traffic conditions and topography. Most of the external factors can be reasonably well estimated using well known engineering formulas, however; the VSL also depends on the pattern of acceleration/deceleration that takes place along the branch. Driving pattern effects depend largely on complex interactions between the driver, the road, the powertrain and traffic. One approach to predicting these complex factors is by classifying road, traffic, and driver and drivetrain combinations into load effects. One simple approach is to classify the arcs of the graph into urban or highway. The US Environment Protection Agency (EPA) provides highway mileage and city mileage for all vehicles sold in the US in the last 15 years based on two standard driving cycles. Clearly having just two cycles is inadequate to capture all the detail necessary, but the model is simple and the data is readily available from www.fueleconomy.gov. A more detailed classification is described in Brundell-Freij and Ericsson [2], where a classification system has been developed based on extensive data collected from instrumented vehicles. Four variables relating to the road type were found to be significant; occurrence and density of junctions controlled by traffic lights, speed limit, function of the street and the type of neighborhood. A large effect was attributed to the power-weight ratio of the vehicle, which presumably is descriptive of the drivers that choose a vehicle with a specific power-weight ratio. A limitation of using instrumented vehicles is collecting adequate data. In the Brundell-Freij and Ericsson study much of the data was collected in Lund, Sweden where there is little topography so road gradient is not considered, although it is well known that road gradient is an important factor where there is topographic relief. Also road design, traffic control policies may vary considerably between political jurisdictions, as may the mix of vehicles. One way to overcome these difficulties is to

use modeling rather that collected data to determine the significant factors for classification. Modeling tools such as the Powertrain System Analysis Toolkit (PSAT) [3] or The MathWorks Simulink/SimDriveLine [4] can be used to estimate the energy or fuel along a route given vehicle design parameters, external load factors such as road gradient and the driver's torque demand along the route. Vehicle design parameters can be obtained from vehicle manufacturers and external factors from published maps. Driver's torque demand involves the psychophysics of driving and can be simulated using software such as VISSIM [5] or MISSION [6]. Vehicles in the fleet may be instrumented to collect actual fuel economy data along the branches they travel. The data may be recovered from the vehicle at a download site and stored in a database. Periodically the database may be used to automatically refine the costs assigned to a class of road segments, and to reclassify segments as needed. The refining of the costs can be done using statistical process control on the existing cost functions, and cluster analysis to discover segments that are outliers in their class and reclassify them to make the cost function more representative.

Vehicle Routing to Minimize Fuel Consumption (a number of formulations for this new variation of the vehicle routing problem. We also outline a solution approach based on composite variable modeling)
The Vehicle Routing Problem to Minimize Mixed-Fleet Fuel Consumption and Environmental Impact (VRPMF) belongs to the class of heterogeneous fleet vehicle routing problems (HVRP). [7] provides a comprehensive classification and review of the main approaches proposed for VRP with a heterogeneous fleet. Specifically, the problem being considered in the paper represents a variant of Heterogeneous VRP with Vehicle-Dependent Routing Costs (HVRPD). They note that solution approaches to this difficult family of problems, both in the literature and in commercial applications, have predominantly been heuristic in nature. These are typically adaptations or extensions of solution techniques for traditional VRP and VRP with Time Windows. In order to capture the complexities (and, in particular, the non-linearities) of VRPMF, we instead propose to leverage the use of composite variable modeling (CVM) to capture the complex real-world details associated with accurately modeling the fuel cost (and associated emissions) of a prescribed route. The idea behind CVM [8] is to embed modeling complexity into the variable definition rather than capturing it explicitly in a model which may then become intractable. For example, in VRPMF, explicitly modeling the cost functions described in Section 2 within the framework of a traditional VRP would make an already difficult problem unsolvable. However, it is far easier to calculate the cost of a given route (for a given vehicle) off-line. We can then formulate a master problem in which each variables represents the assignment of a specific route to a specific vehicle. For a given route, we can compute the total fuel consumption for a given vehicle, and thus the total cost is just the sum of the chosen assignments. Similarly, a route pre-specifies all the customer demands that it meets,

and thus we only need two sets of constraints. The first ensures that each vehicle is assigned to at most route and the second ensures that each customer demand is met exactly once. The challenge, then, is to address the exponentially large number of potential variables. Clearly not all feasible routes can be generated and their costs generated, and even if they could, it would not be possible to solve the resulting exponentially large set partitioning problem. Instead, column generation techniques (originally developed as part of Dantzig-Wolfe Decomposition) can be employed. The idea behind column generation for solving a linear program with an exponential number of variables is to identify candidate pivot variables for the simplex method not by pricing each variables reduced cost directly, but rather by solving a secondary optimization problem (often called a sub problem) which seeks the feasible variable with the most negative reduced cost. If this yields a negative reduced cost variable, then the simplex pivot occurs and the algorithm proceeds. If the most negative reduced cost variable is strictly non-negative, then a certificate of optimality is achieved and the algorithm terminates with a provably-optimal solution. In Dantzig-Wolfe Decomposition, the inherent structure of problem leads to a sub-problem that is pre-defined. Furthermore, it is itself a linear program and thus straightforward to solve. In CVM, because the variable definition is chosen specifically to overcome challenges of a traditional formulation, the sub problem reflects these challenges. Perhaps the most closely related work to our propose approach is that of Taillard [9], who used a heuristic based on column generation techniques to solve HVRPD. Specifically, a large set of candidate routes were generated by solving separate homogeneous VRP problems for each fleet type. The final routes were then selected using a set partitioning formulation to ensure that all demands were met. A key difference between our proposed approach and typical network-based routing problems is that the cost of a route cannot be computed simply by adding the individual arc costs (if so, the sub-problem would be a simple minimum cost flow problem). At first glance, it seems possible to formulate the sub problem as a network flow problem, where each node represents a customer or depot, each arc represents the driving from one node to another, and the cost associated with each arc can fully capture (based on off-line calculations) the cost of this driving. This is not quite true, however the cost on each arc is not independent of the other arcs that are chosen for an individual vehicles route. This is because the fuel consumption on an arc depends on the starting conditions of the vehicle at the first node. If the battery is fully charged, it may be possible to complete most of the driving without relying on gasoline, and the resulting cost and emissions will be low, whereas if the battery is depleted, the cost of the arc will be much higher. Therefore, we must develop a more sophisticated approach to solving the subproblem. Specifically, we propose to solve a variation of the multi-label shortest path problem , an algorithm that is similar to Dijkstras shortest path algorithm, but with an added layer of complexity. Specifically, we must check multiple metrics (not just cost) to determine whether a partial path can be pruned from consideration. One partial path dominates another only if it is less costly and covers the same amount of demand or more and has the same amount of remaining battery charge or more.

Efficiently solving this sub-problem is the key to successfully solving the master problem. We conclude this section by noting that this approach has the added advantage of allowing the user to trade off between time and solution quality. Specifically, the solution quality continues to improve as each new candidate route is added to the master problem for consideration, but high-quality feasible solutions can nonetheless often be found early in the process. Furthermore, this approach naturally lends itself to a parallel implementation. At the highest level, a separate sub-problem can be solved, in parallel, for each vehicle. Furthermore, these individual sub-problems themselves can leverage a parallel architecture for efficient search.

4 Illustrative Example
This section provides a simplified illustrative example of VRPMF. We consider a fleet of two vehicles: the first is a 2009 Ford Taurus front-wheel drive gasoline engine vehicle and the second is 2010 Ford Escape 4-wheel drive hybrid vehicle. Figure 1 shows estimates of the MPG values and environmental scores as provided by the Environmental Protection Agency (EPA) at www.fueleconomy.gov. Note that the Taurus gets 18 MPG in city driving and 28 MPG in highway driving, while the Escape hybrid gets 27 MPG on the highway and 30 MPG in the city. The environmental impact of each vehicle can be evaluated by its carbon footprint and air pollution score. The carbon footprint measures greenhouse gas emissions (primarily carbon dioxide, CO2) that in turn impact climate change. CO2 emissions are closely linked to fuel consumption, since CO2 is the ultimate end product of burning gasoline. The Air Pollution score represents the amount of health-damaging and smog-forming airborne pollutants (such as carbon monoxide, CO, and oxides of nitrogen, NOx) that the vehicle emits on a scale from 0 (worst) to 10 (best). Note that there is little correlation between fuel consumption and these emissions; emissions primarily depend on the emission control technology. Taurus has an Air Pollution score of 6 and Escape Hybrid has a score of 8. Suppose that we have eight customers distributed within a given geographical area as shown in Figure 2. The customers are labeled 1 to 8, while 0 is the depot. The driving distance between the depot and each of the customer sites is presented in Table 1. For this example, we assume that the distance data is symmetrical and shown as X+Y, where X is the number of highway miles and Y is the number of city miles. Further, we assume that each service call requires approximately 90 minutes and that each service agent has to finish her/her route, starting and ending at the depot, within 8 hours. We ignore breaks, and assume that there is no capacity limit on the routes other than the time limit. We begin by solving the traditional problem of minimizing the total fleet vehicle miles traveled (VMT). [In this small problem, this can be done by explicit enumeration.] The total optimum VMT is 97.2. One vehicle visits customers 1, 2, 3, and 4 with a total travel distance of 47.3 miles, comprised of 22.9 city miles and 24.4 highway miles. The other vehicle visits customers 5, 6, 7, and 8 with a total travel distance of 49.9 miles, comprised of 21.9 city miles and 28 highway miles. The

estimated travel time for the first route is 75 minutes and for the second route is 78 minutes. Note that in this variation of the problem, we do not differentiate between vehicles, as we are simply minimizing distance traveled.

Fig. 1. Fleet Composition. Fig. 2. Vehicle Routing Problem

0 0 1 2 3 4 5 6 7 8
14.5+ .9 19+ 1 0+ 11.6 0+ 9.8 0+ 13.6 12.5+ 10 15.5+ 1.1

1
14.5+ .9

2
19+ 1 4.2+ 1.5

3
23.9+ 2.4 10.2+ 1.8 5.7+ 2.2

4
0+ 11.6 0+ 6.3 2.5+ 9.1 0+ 6.7

5
0+ 9.8 0+ 10.3

6
0+ 13.6 0+ 12.1

7
12.5+ 10 12.2+ 2.6 7.7+ 2.9 0+ 3.8 17+ 7.5 0+ 8.6 0+ 4.7

8
15.5+ 1.1 0+ 17.9 20.2+ 2 14.3+ 1.5 0+ 11.6 0+ 7.7 7.4+ 4.1 12.5+ 2.5

4.2+ 1.5 5.7+ 2.2 2.5+ 9.1 4.2+ 11.6 11.4+ 4.7 7.7+ 2.9 20.2+ 2 0+ 6.3 0+ 10.3 0+ 12.1 12.2+ 2.6 0+ 17.9

4.2+ 11.6 11.4+ 4.7 0+ 10.6 0+ 4 0+ 6.5 14+ 6 0+ 3.8 0+ 3.8 0+ 8.6 0+ 7.7 0+ 4.7 7.4+ 4.1

23.9+ 2.4 10.2+ 1.8

0+ 6.7 0+ 10.6 0+ 6.5 0+ 3.8 14.3+ 1.5 0+ 4 14+ 6 17+ 7.5 0+ 11.6

12.5+ 2.5

Table 1. Distance between depot and between customer's sites.

For illustrative purposes, to estimate fuel consumption of each vehicle along the given routes we assume the estimated highway and city MPGs as defined in Figure 1. [Of course, real-world calculations are more complex as we discussed in section 2.] As a result, the fuel consumption of the Taurus for the first route is 2.14 gallons and

for the second route is 2.22 gallons. For the Escape Hybrid, the first route consumes 1.67 gallons and route the second route consumes 1.77 gallons. The optimal solution is to assign the Taurus to the first route, resulting in the consumption of 2.14 gallons, and to assign the Escape to the second route, with an estimated consumption of 1.67 gallons of gasoline. The total fleet fuel consumption for the given solution is 3.81 gallons of gasoline. The results are presented graphically in Figures 3 and 4.

Fig. 3. Solution 1 Graph

Fig. 4. Solution 1 Map

We next reformulate the problem to take into fuel economy, i.e. distinguishing in our optimization between highway and city driving, and treating the two vehicles separately, in recognition of their distinct characteristics. The optimal solution for this problem is presented in Figures 5 and 6.

Fig. 5. . Solution 2 Graph

Fig. 6. Solution 2 Map

Although the VMT increases (from 97.2 to 101.7), the fuel consumption decreases from 3.81 to 3.72 gallons. Note that the optimal routes have changed (see

Figure 4). The route served by the Taurus covers customers 1, 2, 7, and 8 with a total length of 63.3 miles, but of this 54.4 miles is highway driving and only 8.9 miles is city driving. Conversely, the Escape Hybrid is assigned to a route that serves customers 5, 6, 3, and 4, using only city driving (for a total length of 38.4 miles). Table 2 provides a comparative summary of the two solutions. In addition to the economical impact, this demonstrates how a reduction in fuel consumption also leads to a reduction in environmental impact from the fleet operations. First of all, the reduction of fuel consumption reduces the CO2 emissions. In addition, the second solution shifts the time the Taurus spends in the city routes to the Escape Hybrid. As is shown in Figure 1, the Taurus has a lower EPA Air Pollution Score than the Escape Hybrid. Consequently the second solution also reduces health-damaging and smogforming airborne pollutants along populated areas. We could also capture this explicitly in the objective function of our formulation, either by specifying constraints on the total Air Pollution Score (possibly weighted by the location of the route arcs), or by introducing weights in the objective function.

Table 2. Comparative summary of two VRT solutions

5 Conclusion
In this paper we discuss the importance of the heterogeneous fleet vehicle routing problem based on fuel consumption rather than just distance traveled. We describe the complex function that determines how much fuel a given route consumes, and argue that distance is an inadequate surrogate when multiple fleet types, especially varying across different technologies, are used. We provide a simple example to

illustrate how minimizing total miles traveled can yield a very different solution than minimizing fuel consumption. We also discuss solution techniques, specifically based on the use of composite variable modeling, to solve this computationally challenging problem. In the future, we propose to consider emissions as well as explicit fuel costs (which implicitly capture CO2 emissions but not NOx). We also suggest extending VRPMF to include variations and extensions such as those studied in the basic VRP, such as balancing routes, satisfying time windows, etc.

References
1. 2. Chan, C.C. and Chau, K.T.; Modern Electric Vehicle Technology; page 230; Oxford University Press (Nov. 15, 2001); Brundell-Freij, K. and Ericsson, E.; Influence of street characteristics, driver category and car performance on urban driving patterns, Transportation Research Part D 10 (2005), pp. 213-229.
3. 4. PSAT Overview, January 2008,

5.
6.

7.

8.

http://www.transportation.anl.gov/pdfs/HV/412.pdf Rose-Hulman Institute of Technology; Rose-Hulman Institute of Technology Students Design Hybrid Vehicle Powertrain with Simulink and SimDriveLine; 2005; http://www.rosehulman.edu/challengex/downloads/Rose-Hulman_user_story.pdf PTV AG, VISSIM State-of-the-Art Multi-Modal Simulation; http://www.ptvag.com/fileadmin/files_ptvag.com/download/traffic/Brosc hures_Flyer/VISSIM/VISSIM_Brochure_e_2009_HiRes.pdf R. Wiedemann, U. Reiter, Wiedemann Model-Microscopic Traffic Simulation The Simulation System Mission; 1991; http://www.ptvamerica.com/fileadmin/files_ptvamerica.com/library/1970 s%20Wiedemann%20VISSIM%20car%20following.pdf Roberto Baldacci, Maria Battarra, and Daniele Vigo, Routing a Heterogeneous Fleet of Vehicles in B. Golden, et al. (Eds), The Vehicle Routing Problem: Latest Advances and New Challenges, Springer 2008, pp. 3-27. Barlatt A.,Cohn A.,FradkinY., Gusikhin O.; Morford C., Using composite variable modeling to achieve realism and tractability in production planning: An example from automotive stamping, IIE Transactions, Volume
41, Issue 5 May 2009 , pp. 421 436.

9.

Taillard E.D.,A heuristic column generation method for the heterogeneous VRP, Operations Research -- Recherche oprationnelle 33 (1), 1999, pp. 1-14. (Publication CRT-96-03, Centre de recherche sur les transports, Universit de Montral, 1996. Revision 11.2005)